Mechanistic principles of an ultra-long bovine CDR reveal strategies for antibody design

The stalk is important for the stability of antibodies with ultra-long CDR-H3s

We were interested to dissect the design principles of naturally occurring ultra-long CDR-H3s. Specifically, we aimed to understand the importance of the stalk and knob regions. We therefore created NC-Cow1 mutants with a knob deletion (Δknob) or a stalk deletion (Δstalk) (Fig. 1a). We also replaced the amino acids in the stalk of NC-Cow1 wildtype (wt) with glycine residues to obtain a mutant with a flexible stalk (G-stalk) (Fig. 1a). The rationale for replacing the stalk with glycine residues was to determine whether the length of the stalk or the specific interactions between the stalk and the framework were important.

We expressed these NC-Cow1 variants transiently as Fab fragments in Expi293 cells. Immunoprecipitation of the respective supernatants showed that the NC-Cow1 wt and the Δknob mutant were secreted well (Fig. 1c); in contrast, the Δstalk and G-stalk mutants were secreted poorly (Fig. 1c). These results suggest that the stalk is a key element for the organization of the ultra-long CDRs and that not only the presence of a stretch of amino acids is important but a stalk with intrinsic structural features conveying stability and a specific three-dimensional structure.

We purified these proteins to homogeneity and determined their structural and functional characteristics. SEC MALS analyses revealed that all four NC-Cow1 Fab variants are heterodimers with the expected molecular mass (Fig. 1d). The secondary structure of the NC-Cow1 mutants investigated was similar to that of the wt as deduced from FUV CD spectroscopy (Fig. 1e). The thermal stability of the constructs was determined by following the change in CD signal with increasing temperature. Interestingly, the Δknob mutant has an almost identical thermal stability (T_M = 67.5 °C) compared to NC-Cow1 wt (T_M = 65.9 °C), but the Δstalk and G-stalk mutants unfold at lower temperatures with T_M values of 56.3 °C and 53.3 °C, respectively (Fig. 1f; Table 1).

To test how the architecture of the ultra-long CDR affects antigen binding, the wt and variants were assayed concerning their affinity for the HIV-1 Env protein by SPR. As expected, the Δknob mutant did not bind to the antigen with detectable affinity (Fig. 1g). Interestingly, the NC-Cow1 mutants without a stalk or with a stalk made of glycines bound to the HIV-1 Env protein. The affinity constants determined for the constructs (Supplementary Fig. 1) revealed that NC-Cow1 wt binds with a K_D = 3.2 nM; the G-stalk mutant has a similar K_D of 4.3 nM, while the Δstalk mutant showed only a slightly weaker binding and a K_D = 11 nM (Table 1).

Molecular dynamics (MD) simulations on the Fv of NC-Cow1 provided insight into the importance of the stalk. The mutant with deleted stalk could not be modeled due to severe steric clashes between the knob and the Fv framework when the stalk was removed, which corresponds to the lower stability and antigen-binding affinity of the Δstalk mutant compared to wt. MD simulations on the G-stalk mutant indicate that replacing the stalk with Gly residues strongly increases the motion of the knob relative to the Fv framework (magenta lines in Fig. 1h and i); however, the root-mean-square deviation (RMSD) of the knob itself with respect to the start knob structure is remarkably similar between wt and G-stalk (orange lines in Fig. 1h, i). Hence, the knob retains its rigidity when the stalk is highly flexible, which explains the preserved antigen binding by the G-stalk mutant. The RMSD of the core part of the G-stalk (blue lines in Fig. 1h and i) is on average slightly larger (<RMSD> ~2.2 Å) than the corresponding mean deviation for the wt (<RMSD> ~1.5 Å). Hence, the highly flexible stalk disturbs the Fv framework structure of NC-Cow1 which agrees with the lower thermal stability of G-stalk compared to wt.

Taken together, the data reveals that the specific stalk structure in NC-Cow1 wt is important for the overall stability but less for the folding of the knob.

Stalk extensions in the bovine ultra-long CDR-H3 are not beneficial

Knowing that stalk deletions have detrimental effects on the stability of NC-Cow1, we asked how stalk extensions will affect the antibody. Parts of the two strands of the stalk in NC-Cow1 form a β-ribbon (Fig. 1b)¹⁸. We were interested whether this naturally-occurring stalk could be extended by replicating the rigid β-ribbon structure. We therefore repeated these complementary residues to obtain mutants with a double and a triple stalk, named (stalk)₂ and (stalk)_3, respectively (Supplementary Fig. 2a). We observed that extending the stalk in this way reduced the secretion levels of the corresponding Fab fragments from mammalian cells (Supplementary Fig. 2b). The secondary structures of the purified variants remained unchanged (Supplementary Fig. 2c), but the thermal stability of the Fabs with longer stalks was lower compared to the wt (Table 1 and Supplementary Fig. 2d). The (stalk)₂ and (stalk)₃ mutants bound to the antigen with K_Ds of 5.7 nM and 5.3 nM, respectively (Table 1 and Supplementary Fig. 2e and f).

Taken together, the stalk deletions and extensions show that while both shortening and elongating the stalk is possible, it is not beneficial. Thus the stalk length in NC-Cow1 wt has evolved for optimal folding, stability and binding to the antigen.

Replacing the ultra-long CDR-H3 with short bovine CDRs reduces Fab stability

Bovine antibodies like NC-Cow1 have the longest CDR-H3s found in nature. One could presume that such long CDRs are avoided by evolution because they could impose constraints on folding and stability. However, despite the complexity of its CDR-H3, the NC-Cow1 Fab was well secreted (Fig. 1c) and thermally stable (Fig. 1f). This raises the question whether the framework in these bovine antibodies has exceptional stability that can compensate for potential destabilization induced by the ultra-long CDR-H3. To test this hypothesis, we replaced the ultra-long CDR-H3 in NC-Cow1 with short bovine CDR-H3 loops from the bovine antibodies BS1, BS2, B4 and B13 (Supplementary Fig. 3a)^13,19. However, in contrast to our assumption, the mutants with short bovine CDR-H3s were secreted poorly compared to the NC-Cow1 wt (Supplementary Fig. 3b). In line with the potential folding problems in the cell, the grafted variants with short bovine CDR-H3s affected the secondary structure of the Fabs (Supplementary Fig. 3c) and reduced their thermal stabilities (Table 1 and Supplementary Fig. 3d). We wondered whether the high stability is a specific feature of the cow Fabs with ultra-long CDRs and therefore also characterized bovine Fabs which naturally contain short CDR-H3s, namely B4 and B13. We found that they formed heterodimers (Supplementary Fig. 3e), had the characteristic secondary structures (Supplementary Fig. 3f), and high thermal stability (Supplementary Fig. 3g).

Thus, the bovine antibodies with ultra-long CDR-H3s evolved to maintain the typical stability of bovine antibodies and are in this respect indiscriminable from their natural counterparts with short CDR-H3s. Surprisingly, introducing short loops instead of the ultra-long CDR-H3 in the framework is detrimental for stability.

The disulfide bonds rigidify the knob to allow antigen binding

The bovine ultra-long CDR-H3s contain several cysteines that form disulfide bonds¹³; NC-Cow1 has three disulfide bonds that form a characteristic pattern in the knob (Fig. 2a)¹⁸. Alanine scanning had indicated that some of these cysteines (Cys II and Cys V in Fig. 2a) are crucial for binding to the antigen¹⁸. However, little is known of the importance of each disulfide bond for NC-Cow1 stability and antigen binding.

To study the importance of disulfide bonds in the knob of NC-Cow1, we created mutants where each pair of cysteines that forms a disulfide bond (Fig. 2a) was replaced with a pair of serines, (C(I,IV)S, C(III,VI)S, and C(II,V)S). In addition, we created a mutant, C(I-V)S, where all six cysteines in the knob were replaced by serine residues. All C → S mutants were secreted at levels comparable to those of the NC-Cow1 wt Fab in mammalian cells (Fig. 2b) and exhibited similar CD spectra (Fig. 2c); remarkably, however, all mutants had lost their antigen-binding properties (Fig. 2d). The question remaining was whether the disulfide bond deletions in the knob affect the overall stability of NC-Cow1 Fab. Preliminary experiments monitoring the thermal stability of the C → S mutants by FUV CD spectroscopy revealed transitions similar to those observed for NC-Cow1 wt (Supplementary Fig. 4a). To determine whether there are subtle changes in the thermal unfolding that were not detected by FUV CD, we used differential scanning calorimetry (DSC). NC-Cow1 wt Fab exhibited one unfolding peak in the thermogram with a T_M of 69.4 °C (Fig. 2e) that corresponds well to the observed transition in FUV CD (Fig. 1f). The mutants with pairwise C → S substitutions also showed only one unfolding transition and had only slightly lower T_M values (Fig. 2e and Table 1). Noteworthy, even the C(I-V)S mutant that cannot form any disulfide bonds in the knob exhibits a thermal stability very similar to the NC-Cow1 wt with a T_M of 67.1 °C.

Since antigen binding is abolished in the cysteine mutants, we hypothesized that removing disulfide bonds from the knob will cause structural changes that could affect not only the interaction with the antigen but also the hydrodynamic properties of NC-Cow1 Fab. To test this, we performed SEC MALS experiments and observed that all C → S mutants exist as heterodimers with the correct molecular mass (Fig. 2f); however, the mutants with replaced pairs of cysteines eluted earlier compared to the wt and showed asymmetric peaks indicating conformational heterogeneity (Fig. 2f). The C(I-V)S mutant with no cysteines in the knob showed the shortest retention time (Fig. 2f).

We then used MD simulations to explain the effects of disulfide bond removal on antigen binding and stability. Comparative MD simulations on the Fv of the NC-Cow1 wt and the cysteine variants indicate that the framework is not affected as no deviation (RMSD) was observed when disulfide bonds are removed from the knob (Fig. 2g). This result is in line with the observed small effect of the disulfide bond removal on protein stability. Also, the overall motion of framework plus knob is only modestly increased upon disulfide bond removal (Fig. 2h). However, the RMSD of the knob itself increased significantly over time already upon the removal of one disulfide and even larger deviations relative to the starting structure were observed when removing all disulfide bonds (Fig. 2i). These structural changes in the knob of the ultra-long CDR-H3 in the absence of disulfide bonds explains the experimentally observed loss of antigen binding of these variants.

Because the formation of multiple disulfide bonds within one structural element can be challenging, we wondered whether a salt bridge can be introduced to replace a disulfide bond. To test this, we created a mutant where we replaced Cys I and Cys IV (Fig. 2a) with Arg and Glu respectively to create a mutant called NC-Cow1 SB. This mutant was well secreted but did not bind to the antigen (Supplementary Fig. 4).

Overall, the C → S mutants of NC-Cow1 revealed important information. The disulfide bonds in the knob are crucial for antigen binding by imposing conformational restraints in the paratope; however, the disulfide bonds in the knob do not contribute significantly to the overall Fab stability – even when disulfide bonds are absent from the knob of NC-Cow1. From a stability perspective, our results show that NC-Cow1 can tolerate not only folded and rigid knobs in its ultra-long CDR-H3 but also a large, disordered domain.

The bovine knob can be replaced by long human CDRs

It had previously been shown that the knob in bovine antibodies with ultra-long CDR-H3 can be replaced by short peptides^20,21 or small cytokines^22,23 to create functional fusion proteins. Since the NC-Cow1 Fab remained stable after we removed all disulfide bonds in the knob, we wondered whether the structure would tolerate the replacement of the knob with non-bovine long CDRs. In fact, there are several human antibodies that have a long CDR-H3 that folds into a hammerhead structure (e.g. the PG16 antibody)²⁴ or contains a beta strand (e.g. the VRC26.25 antibody)²⁵ that are important for binding to the antigen. The CDR-H3 of PG16 does not contain cysteines, while the CDR-H3 of VRC26.25 has one disulfide bond. We were therefore interested if the grafting of these two different human CDR structures will affect the stability of NC-Cow1. Moreover, we wanted to test whether the grafted human CDR-H3 will still exhibit binding to the antigen. We therefore grafted sequences from the CDR-H3 of PG16 and VRC26.25 in the place of the knob in NC-Cow1 to create the wt→PG16 and wt→VRC26.25 mutants (Supplementary Fig. 5a). These NC-Cow1 Fab mutants were well secreted (Supplementary Fig. 5b) and had wildtype-like secondary structure (Supplementary Fig. 5c). The thermal stability of the mutants with replaced knobs was only slightly lower (Table 1 and Supplementary Fig. 5d); wt→PG16 has a T_M of 65.3 °C, the T_M of wt→VRC26.25 is 67.9 °C. Like the C → S mutants in the previous section, wt→PG16 and wt→VRC26.25 showed differences in the peak shape and elution time in SEC MALS (Supplementary Fig. 5e) and did not bind to the antigen (Supplementary Fig. 5f).

All in all, it seems that the human CDR-H3 sequences can replace the knob of NC-Cow1 with only minor effects on folding and thermal stability. These findings suggest that the landscape of possible donor sequences that can be grafted onto ultra-long CDR-H3s is not limited to short peptides and compactly folded proteins. However, the grafted sequences need to contain the entire antigen-binding properties, like the bovine knob, to obtain functional antibodies.

Bovine antibodies with ultra-long CDR-H3 can be humanized by two approaches

NC-Cow1 broadly neutralizes HIV and presents an attractive option for the prevention and therapy of acquired immunodeficiency syndrome¹⁵. For clinical success, animal antibodies need to be humanized to reduce immunogenicity risks²⁶. The humanization of bovine antibodies with ultra-long CDR-H3s has not been demonstrated yet. It remained unclear whether human antibody scaffolds can tolerate the ultra-long bovine CDR-H3 without detrimental effects on folding, stability, and antigen binding.

Humanization has been achieved by replacing the CDRs in a human scaffold with the CDRs from an animal antibody²⁷. This approach often requires the grafting of several or all 6 CDRs that form the antigen-binding site. In bovine antibodies with ultra-long CDR-H3, the antigen-binding site is formed by the knob; however, it was suggested that successful humanization might require grafting of the remaining CDRs that could provide structural support for the ultra-long CDR-H3⁷.

To test whether the humanization of bovine antibodies with ultra-long CDR-H3 is possible, we tested three different human scaffolds (Fig. 3a). First, we selected an intrinsically stable Fab scaffold (trastuzumab) that has already been used in humanization²⁸ and maintained excellent stability²⁹. Second, we used BLAST³⁰ to compare bovine and human antibody sequences in the PDB. We identified several human Fd sequences with high homology to bovine antibodies (Supplementary Fig. 6) and subsequently the corresponding human and bovine LCs (Supplementary Fig. 7). We selected the human Fab 32H+109L³¹ that has overall high homology to the bovine Fab fragments with ultra-long CDR-H3s as a human scaffold for our experiments. Third, we included a human Fab (PGT145)³² with a long (31 residue) CDR-H3. The sequence alignments of these proteins and the corresponding mutants can be found in supplementary data.

We replaced the CDRs in the human Fabs with all 6 CDRs from NC-Cow1 (Fig. 3b). The CDR grafting reduced the production yields of the trastuzumab Fab and PGT145 Fab 20 times and 10 times, respectively (Table 1). Interestingly, the yields of 32H+109L were not negatively affected by the CDR grafting. All three human scaffolds with CDRs from NC-Cow1 bound to the HIV-1 Env antigen with K_Ds between 3 nM and 4.9 nM (Table 1 and Fig. 3c and Supplementary Fig. 8a). Due to low yields of some purified proteins, we used fluorescence spectroscopy (nanoDSF) to study the thermal stability of the chimeric constructs and observed that all chimeric Fabs were less stable compared to the wt proteins (Table 1 and Supplementary Fig. 8d). DSC analysis of the best acceptor (32H+109L) of the cow CDRs (Fig. 3d) confirmed the T_Ms determined by the fluorescence-based approach (Table 1).

To reduce the size of the transferred structural element, we inserted only the knob (without the stalk) of NC-Cow1 into the tip of human CDR-H3s (Fig. 3e). Overall, the knob grafting had less negative impact on folding and stability compared to grafting all CDRs; the production yield of trastuzumab+knob was the same as observed for the wt, while the yields of 32H+109L + knob and PGT145 + knob were around 10 and 4 times lower compared to the wt (Table 1).

The K_D of trastuzumab+knob to the HIV-1 Env protein (Fig. 3f and Supplementary Fig. 8b) was 17.6 nM (Table 1); this was the weakest affinity of all chimeric constructs tested and similar to the value obtained for the Δstalk mutant of NC-Cow1. This suggests that in both trastuzumab+knob and NC-Cow1 Δstalk, the knob is located close to the V_H framework, which has negative effects on antigen binding. The 32H+109L + knob and PGT145 + knob had K_Ds of 6.5 nM and 7.7 nM, respectively (Table 1); the CDR-H3s of these two human Fabs are longer which positions the inserted knob further away from the V_H framework in comparison to trastuzumab. In addition to the SPR measurements, we tested whether the human Fabs with grafted bovine knobs bind to the soluble HIV Env trimer (Supplementary Fig. 8c). All three Fabs bound the trimeric HIV Env as seen by the molecular weight shift in SEC-MALS. It turned out that 32H+109L + knob and PGT145 + knob are better binders than trastuzumab + knob which supports our SPR data. In agreement with the results shown above, the optimal distance between the V_H and the knob seems to be important.

Knob grafting had less detrimental effects on the thermal stability of the Fabs compared to grafting all 6 CDRs (Table 1 and Supplementary Fig. 8d); correspondingly, the secondary structure of the human Fabs with grafted knobs was more similar to the parent molecule than the human Fabs with 6 grafted bovine CDRs (Fig. 3h). We were also interested whether the knob grafting results in more aggregation-prone molecules compared to the parent antibodies. We therefore used SEC to analyze purified wt and knob mutant pairs (Supplementary Fig. 9a). There were only slightly more aggregates detected in the chimeric proteins and none of the Fabs aggregated during incubation at 50 °C for 24 h (Supplementary Fig. 9a). Since knob grafting is an unconventional humanization approach, we also wanted to test whether we can produce 32H+109L + knob and PGT145 + knob as full-length IgGs. We therefore fused the Fd sequences to the Fc from trastuzumab to obtain HCs and produced the IgGs in mammalian cells. We purified the IgGs secreted in the cell supernatant by protein A chromatography and analyzed the proteins with SEC MALS (Supplementary Fig. 9b). The chimeric IgGs with a grafted bovine knob showed one main peak with a molecular mass around 150 kDa and contained only small fractions of aggregates and fragments (Supplementary Fig. 9b).

Further, we were interested whether the motion of the bovine knob relative to the Fv framework is preserved upon grafting the bovine CDRs or only the bovine knob onto human antibodies. We therefore performed MD simulations with PGT145 wt, PGT145 with all 6 CDRs from NC-Cow1, and with PGT145 + knob (Supplementary Fig. 10). We observed that the overall knob dynamics in the chimeric structures was similar to that of the NC-Cow1 wt. Thus, it can be expected that also the positioning of the knob and its structural isolation in the chimeric constructs with a grafted bovine knob is similar to NC-Cow1 wt.

Overall, both humanization approaches (CDR grafting or knob grafting) seem to work for the bovine NC-Cow1. In the case of CDR grafting, the human scaffold with the highest homology to the bovine sequences (32H+109L) proved to be the best acceptor yielding a well-secreted chimeric protein combining excellent antigen binding with favorable biophysical properties. In the case of knob grafting, a human Fab with a longer CDR-H3 should be used for optimal antigen binding.

Antigen-binding peptides from de novo design can be grafted as ultra-long CDRs

As demonstrated above, each of the disulfide bonds in the knob is crucial for antigen binding. This restricts approaches for the screening of display libraries to obtain binders; e.g. just 17% of a mammalian library with cysteine-rich peptides was shown to be properly folded³³. Thus, it would be advantageous to combine the stalk concept with different structural elements displaying the desired antigen-binding properties which are ideally free of cysteines.

Recently, cysteine-free tri-helical peptides that bind with picomolar affinity to the spike protein of SARS-CoV-2 have been identified by de novo design³⁴. One of the most potent peptides (LCB1) isolated by this approach interacts with the receptor binding domain (RBD) of SARS-CoV-2 through two alpha helices (marked in green in Fig. 4a). We hypothesized that we could use these helices in the context of an ultra-long CDR-H3. To test this, we inserted the sequence encoding the two helices into the CDR-H3 of three human and one bovine Fab fragment (Fig. 4a). All four Fab fragments with the LCB1 insertion were correctly folded (Supplementary Fig. 11a) and bound with sub-nanomolar affinity to the RBD of SARS-CoV-2 (Fig. 4b, c and Supplementary Fig. 11b). Remarkably, the Fab-LCB1s formed complexes with molecular masses corresponding to that of two Fab fragments (Fig. 4d). Thus, the bi-helical insertion from LCB1 seems to provide not only a high-affinity binding site for SARS-CoV-2, but also a dimerization motif.

We then used SEC MALS to determine the stoichiometry of the complex between SARS-CoV-2 RBD and a dimeric Fab NC-Cow1-LCB1. When mixing the Fab with a molar excess (5x) of the RBD, the Fab-LCB1 complex had the molecular weight of NC-Cow1-LCB1 bound to two RBDs (Fig. 4e). Thus, one dimeric Fab-LCB1 can bind two RBDs.

We also tested sub stoichiometric ratios between the RBD and NC-Cow1-LCB1 (Fig. 4f). When the [RBD]/[NC-Cow1-LCB1] ratio was 0.5 or 1.0, mostly a complex with only one RBD was formed; at a ratio of 1.5, both complexes with one and two RBD bound to the NC-Cow1-LCB1 were observed; at a ratio of 2.0, the dominant species was NC-Cow1-LCB1 bound to two RBDs.

An important question for therapeutic application is whether the engineered Fabs can block the interaction of the SARS-CoV-2 RBD with the human ACE2 protein. The RBD alone binds efficiently to the immobilized ACE2 receptor in biolayer interferometry (Fig. 4g). When the RBD is premixed 1:1 with NC-Cow1-LCB1, the binding to ACE2 is abolished. Thus, the Fabs with LCB1 peptide insertion block the interaction between the RBD and the ACE2 receptor.

To test whether the constructs neutralize infectious virus and prevent cell infection, we determined the neutralization efficiency of the four Fabs with LCB1 insertions for SARS-CoV-2, isolated early in the pandemic in January 2020, and the SARS-CoV-2 Alpha (B.1.1.7) variant of concern which is currently one of the most prevalent variants worldwide. All four proteins neutralized SARS-CoV-2 very potently with 50% inhibitory concentrations (IC50) between 5.7 nM and 7.9 nM (Fig. 4h). The Alpha variant was also neutralized efficiently with IC50 values between 27 nM and 47 nM.

Interestingly, the Fab-LCB1 proteins have similar thermal stabilities as the wt Fabs and the Fabs with a bovine knob (Table 1 and Supplementary Fig. 11c). This is remarkable since the structure of the mini-domain from LCB1 and the bovine knob are completely different. It reveals that the same Fab scaffold can equally well tolerate the insertion of different naturally occurring and de novo designed structures in the CDR-H3. This finding reveals a new strategy for the rapid discovery of highly potent antibodies for biomedical applications.